AI agents can automate useful work, but full autonomy is rarely the right starting point. Many production actions still need human judgment, approval, or later audit because the cost of a mistake is high or the system does not have enough context. Human-in-the-loop (HITL) design makes those control points explicit.

Human-in-the-loop systems combine automation with clear places for people to step in. The agent handles routine work. Humans review important outputs, approve risky actions, resolve ambiguity, and correct mistakes. Those corrections can later become evaluation data, examples, or training data.

This chapter covers approval workflows, confidence-based escalation, asynchronous review queues, correction loops, autonomy levels, and UX patterns that make human review useful instead of a checkbox exercise.